Prediction of Age-related Macular Degeneration in the General Population

Abstract

Purpose

Prediction models for age-related macular degeneration (AMD) based on case-control studies have a tendency to overestimate risks. The aim of this study is to develop a prediction model for late AMD based on data from population-based studies.

Design

Three population-based studies: the Rotterdam Study (RS), the Beaver Dam Eye Study (BDES), and the Blue Mountains Eye Study (BMES) from the Three Continent AMD Consortium (3CC).

Participants

People (n = 10106) with gradable fundus photographs, genotype data, and follow-up data without late AMD at baseline.

Methods

Features of AMD were graded on fundus photographs using the 3CC AMD severity scale. Associations with known genetic and environmental AMD risk factors were tested using Cox proportional hazard analysis. In the RS, the prediction of AMD was estimated for multivariate models by area under receiver operating characteristic curves (AUCs). The best model was validated in the BDES and BMES, and associations of variables were re-estimated in the pooled data set. Beta coefficients were used to construct a risk score, and risk of incident late AMD was calculated using Cox proportional hazard analysis. Cumulative incident risks were estimated using Kaplan–Meier product-limit analysis.

Main Outcome Measures

Incident late AMD determined per visit during a median follow-up period of 11.1 years with a total of 4 to 5 visits.

Results

Overall, 363 participants developed incident late AMD, 3378 participants developed early AMD, and 6365 participants remained free of any AMD. The highest AUC was achieved with a model including age, sex, 26 single nucleotide polymorphisms in AMD risk genes, smoking, body mass index, and baseline AMD phenotype. The AUC of this model was 0.88 in the RS, 0.85 in the BDES and BMES at validation, and 0.87 in the pooled analysis. Individuals with low-risk scores had a hazard ratio (HR) of 0.02 (95% confidence interval [CI], 0.01–0.04) to develop late AMD, and individuals with high-risk scores had an HR of 22.0 (95% CI, 15.2–31.8). Cumulative risk of incident late AMD ranged from virtually 0 to more than 65% for those with the highest risk scores.

Conclusions

Our prediction model is robust and distinguishes well between those who will develop late AMD and those who will not. Estimated risks were lower in these population-based studies than in previous case-control studies.

Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly of industrialized countries.^1,2 Approximately 21 million elderly individuals are affected worldwide, and this number is expected to increase dramatically with the aging population.^3,4 Age-related macular degeneration can be divided into several stages: early AMD, characterized by subcellular deposits (drusen) and pigmentary changes, and late AMD, subdivided into atrophy of the retinal pigment epithelium (dry AMD) and choroidal neovascularization (wet AMD). Despite improved treatment options, late AMD can cause irreversible blindness, whereas severe stages of early AMD are mostly asymptomatic but signal a high risk of progression to late AMD.⁵

Age, early AMD phenotype, and genetic and environmental factors play important roles in the pathogenesis of late AMD.^6–11 These factors may be used to predict this end stage and to identify high-risk individuals. Reasons for assessing predictive values may be risk-dependent (personalized) patient care and surveillance strategies for therapy. Future intervention research, such as randomized, controlled clinical trials, can use prediction models to select individuals with a high risk of outcome events.

Previously reported prediction models were based on selections of cases and nonaffected controls.^12–28 Most studies compared only the extreme ends of disease, excluding the majority of the population with an intermediate disease risk. This has inherent methodological concerns because the disease risk is overestimated by design. Population-based studies include a whole spectrum of risk levels, and therefore findings from these studies would be more generalizable²⁹ and better suited for clinical implementation.

In this study, we present a prediction model for late AMD based on population-based cohort studies from 3 continents. We optimized a prediction model in one of the cohorts and subsequently validated it in the other 2 cohorts. We included established genetic, environmental, and clinical risk factors in the model, assessed relative and cumulative risks, and provided a risk score that can be used to estimate the risk of AMD in individuals.

Methods

For this article, we followed the guidelines for genetic risk prediction studies.³⁰

Study Populations

The Three Continent AMD Consortium (3CC) consists of 4 population-based studies: the European Rotterdam Study (RS), the American Beaver Dam Eye Study (BDES), the Los Angeles Latino Eye Study, and the Australian Blue Mountains Eye Study (BMES). For the purposes of this study, the Los Angeles Latino Eye Study was excluded because of the absence of genotype and follow-up data.

The RS is a prospective, population-based cohort study investigating chronic diseases in the elderly. All inhabitants aged 55 years and older living in a suburb of Rotterdam, The Netherlands, were invited to participate in the study.^31,32 Of the initial cohort of 10 275 eligible individuals, 7983 (78% of those eligible) participated in the overall study (98% were white). The ophthalmologic part began later and included 6780 participants (78% of those eligible). Baseline examinations took place from 1990 to 1993, and 4 follow-up examinations were performed in 1993–1995, 1997–1999, 2002–2004, and 2009–2011. The Erasmus Medical Center Ethics Committee approved the study, which complies with the Declaration of Helsinki. All participants gave written informed consent for participation in the study.

The BDES is a prospective cohort study investigating eye diseases among the population of Beaver Dam, Wisconsin.³³ To identify all residents in the city or township of Beaver Dam who were aged 43 to 84 years, a private census was performed from 1987 to 1988. Of the 5924 eligible individuals, 4926 (83% of those eligible) participated in the baseline examination between 1988 and 1990 (99% were white). There were follow-up examinations every 5 years: 1993–1995, 1998–2000, 2003–2005, and 2008–2010. The BDES was approved by the institutional review board from the University of Wisconsin-Madison and adhered to the tenants of the Declaration of Helsinki. All participants provided signed, informed consent for participation in the study.

The BMES is a prospective cohort study of eye diseases and other health outcomes in an urban population.³⁴ All residents aged 49 years or older, living in 2 postcode areas of the Blue Mountains region in West Sydney, Australia, were invited to participate in the study. In 1992–1994, baseline examinations were performed in 3654 participants (82.4% of those eligible). Reexaminations were performed after 5, 10, and 15 years (in 1997–1999, 2002–2004, and 2007–2009, respectively). All BMES examinations were approved by the human research ethics committees of the Western Sydney Area Health Service and the University of Sydney and complied with the Declaration of Helsinki. All participants provided written, informed consent for participation in the study.

Participants were eligible for the current analysis when genotype data, as well as gradable fundus photographs at baseline and at least 1 follow-up eye examination were available (Fig 1, available at http://aaojournal.org). People with late AMD at baseline were excluded. This resulted in 4753 (RS), 3542 (BDES), and 1811 (BMES) participants available for analysis, with a median follow-up of 10.7 years in RS (interquartile range [IQR], 12.8), 15.6 years in BDES (IQR, 10.4), and 11.8 years in BMES (IQR, 5.6). In total, 10 106 participants with a median follow-up of 11.1 years (IQR, 11.0) were included in the analysis. To investigate possible selection bias, we analyzed whether people excluded from this study differed in the baseline level of AMD from those who were included. The 2 groups did not differ in early AMD levels (10–40) after adjustment for age and sex (P = 0.95).

Diagnosis of Age-Related Macular Degeneration

All participants underwent fundus photography after pharmacologic mydriasis. Fundus transparencies of all studies were graded according to the Wisconsin Age-Related Maculopathy Grading^35,36 by trained graders under the supervision of senior retinal specialists or senior researchers (RS: P.T.V.M.dJ., J.R.V., C.C.W.K.; BDES: B.E.K.K., R.K.; BMES: P.M., J.J.W.). The graded fundus photographs were classified using a classification common to all studies: the 3CC AMD severity scale³⁷ (Table 1, available at http://aaojournal.org). All prevalent and incident late AMD cases from each of these 3 studies were cross-checked by investigators of the other 2 studies, with consensus obtained via discussion during multiple teleconferences. The eyes of each participant were graded and classified separately, and the eye with the more severe grade was used to classify the person.

Genotyping

Genomic DNA was extracted from peripheral blood leukocytes. All eligible study participants in the RS were genotyped with the Illumina Infinium II HumanHap550 array or Taqman assays (Applied Biosystems, Foster City, CA). HapMap CEU data (release #22) was used for imputation.

DNA from BDES participants was extracted from the buffy coats of blood obtained at baseline examinations or subsequent examinations that were stored frozen at −80°C. DNA samples arrayed in 96-well plates were submitted for genotyping via an Illumina iSelect Custom Genotyping Panel (Illumina Inc., Hayward, CA) at the Genomics Core Facility at Case Western Reserve University or via the KASP Assay at LCG Genomics (Teddington, Middlesex, UK). The data collected were analyzed using Illumina's Genome Studio or the KASP SNP Genotyping System. The assays were controlled for quality by examining cluster separation values and call frequency. Untyped single nucleotide polymorphisms (SNPs) were imputed using HapMap CEU (release #22) as reference.

In the BMES, all participants with DNA available were genotyped using Illumina Human670-Quad v1 custom array at the Wellcome Trust Centre for Human Genetics, Sanger Institute, Cambridge, UK, as part of the Wellcome Trust Case Control Consortium 2. A smaller subset of participants (N = 1356) was also independently genotyped using the Illumina 610-Quad genotyping array at the Hunter Medical Research Institute, Newcastle, Australia. After quality control, the genotyped data were imputed from the 1000 Genomes (version 1) reference using IMPUTE software.

Selection of Single Nucleotide Polymorphisms

For selection of AMD genes, we reviewed publications on AMD genetics^38,39 and prediction of AMD.^12–28 From these, we selected 41 tag SNPs that were available in all 3 cohorts and that were not in linkage disequilibrium (r² < 0.60). Genotypes of SNPs were coded as 0 for carriers of 2 major alleles, 1 for the heterozygous genotype, and 2 for carriers of 2 minor alleles. When none of the cases were carriers of 2 minor alleles, genotypes of the SNPs were coded as 0 for carriers of 2 major alleles and 1 for all carriers of at least 1 minor allele. As a first step, each SNP was tested for association with late AMD in all 3 cohorts.

For each locus with multiple SNPs, we performed a backward Cox proportional hazard analysis to determine the best predictive SNPs for incident late AMD within each locus with data from the RS.

Assessment of Nongenetic Variables

All nongenetic variables used in the analyses were assessed using baseline data. Information on cigarette smoking was obtained in an interview at baseline and categorized as never, former, and current. Height, weight, and blood pressure were measured at the beginning of baseline examination. Body mass index (BMI) was calculated by dividing weight (kilograms) by the height (meters squared). The BMI variable was categorized as not overweight or obese (BMI<25) and overweight or obese (BMI>25). Age (years) at baseline was categorized in 3 categories; <65, 65–75, and >75 years. Baseline AMD grading was entered into the analysis as categorized variables with levels 10 to 40.

Statistical Analyses

Throughout the entire study, incident late AMD was used as the outcome variable; nonincident late AMD, including those participants remaining at an early AMD stage, was used as the reference group. All analyses were performed using SPSS version 20.0 (SPSS Inc., Chicago, IL).

Variables were analyzed for association with incident late AMD in the 3 cohorts using Cox proportional hazard analysis, adjusting for age and sex. We constructed 5 different models based on a minimal and a maximal selection of clinical, genetic, and environmental factors. Model 1 was a minimal model including only age and sex; model 2 was a nongenetic model including age, sex, and environmental and ocular factors; model 3 was a minimal genetic model including age, sex, and major genetic AMD risk variants (CFH Y402H, ARMS2 A69S, C3 R102G, C2 L9H, CB R32Q); model 4 was a maximal genetic model including age, sex, and 26 genetic AMD risk variants (see “Selection of Single Nucleotide Polymorphisms”); and model 5 was a maximal gene–environmental model including age, sex, environmental and ocular factors, and the 26 genetic AMD risk variants. For each model, we calculated the area under the curve (AUC) of incident late AMD in the RS and validated the best model in the BDES/BMES.

Subsequently, we estimated the association of all variables in the best model using multivariate Cox proportional hazard analysis in the pooled dataset of all 3 cohorts and calculated the AUC. Calibration of the model was tested using the Hosmer–Lemeshow test. This goodness-of-fit test shows how well the predicted risks match the observed risks. To test whether nonmajor genetic AMD risk variants could be discarded from this model without jeopardizing the AUC, backward regression (eliminating SNPs with P > 0.05) using Cox proportional hazard analysis was carried out and the AUC of the new model was calculated within the pooled dataset. In this dataset, we estimated the beta coefficient of each variable from the best model using multivariate Cox proportional hazard analysis. The estimated beta coefficient of a variable was the individual risk score of that variable. Next, we created a summary risk score based on the sum of the beta coefficients from the multivariate Cox proportional hazard analysis. Risk scores were rounded off, and frequencies of the risk scores were plotted stratified for incident late AMD and no AMD. We calculated the risk of incident late AMD with the middle risk score (3) as the reference using Cox proportional hazard analysis. Risk scores at the extreme ends were pooled to increase sample size because of limited numbers.

We calculated the cumulative risk of incident late AMD per risk score. We assigned the age of onset for incident late AMD as the median between the examination at which late AMD was first observed and the previous examination. For participants who did not develop late AMD, we used age at last examination for censoring. All participants aged 90+ years were censored at age 90 years to maintain unbiased estimates. Risks were calculated using Kaplan–Meier product-limit analysis. Participants who died or were lost to follow-up were censored at the time of the last examination. Cumulative risks stratified for the risk score were compared with the overall cumulative risk based on incidence of late AMD (prior risk) using log-rank tests of equality (Mantel–Cox).

Missing data were encountered in the analysis of each model. Only participants with data on all variables in the model entered the analysis.

Results

In total, 363 participants developed incident late AMD during a median follow-up time of 11.1 years (IQR, 11.0), of whom 132 were in the RS (follow-up 10.7 years; IQR, 12.8), 153 were in the BDES (follow-up 15.6 years; IQR, 10.4), and 78 were in the BMES (follow-up 11.8 years; IQR, 5.6). Incidence rates for the 3 studies were 2.89, 2.96, and 3.66 per 1000 person-years for the RS, BDES, and BMES, respectively. The distribution of demographic characteristics and environmental risk factors differed slightly among the 3 cohorts (Table 2). Because the inclusion criteria for age were higher than for the other studies, participants in the RS were older. Early AMD (level 20–40) at baseline was more frequent in RS participants, BMI was higher in BDES participants, and current smoking was less frequent in BMES participants. The frequency of genetic risk alleles was not significantly different among the 3 studies, although there were slight differences in genotype distributions. Visual inspection of a principle component analysis of all genetic data against HapMap CEU data (National Center for Bio-technology Information build 36, release 22) as reference showed similar plots for the 3 cohorts (data not shown).

Table 2. General Characteristics of Participants Included in the Analysis.

Age Inclusion Variables	RS age 55+ (N = 4753),	BDES age 45+ (N = 3542),	BMES age 49+ (N = 1811),
Mean age, yrs (SD)	67.6 (7.8)	60.2 (10.4)	64.0 (8.3)
Age categorized (yrs)
≤65	42.5	67.3	56.8
65–75	39.1	24.0	33.1
75+	18.4	8.7	10.1
Male sex	42.4	42.5	42.8
AMD baseline grade
Level 10	83.8	80.2	87.4
Level 20	9.1	14.0	8.8
Level 30	6.0	5.2	3.4
Level 40	1.1	0.6	0.3
Smoking
Never	33.9	45.1	51.4
Past	43.5	34.9	37.1
Current	22.6	20.0	11.5
Mean BMI, kg/m2 (SD)	26.3 (3.6)	28.8 (5.4)	26.2 (4.2)
BMI categorized
≤25	38.3	24.9	41.7
25+	61.7	75.1	58.3
CFH (Y402H) rs1061170
TT	41.8	39.7	42.5
TC	45.0	47.0	46.5
CC	13.1	13.4	11.0
CFH rs12144939
GG	64.3	64.2	65.1
GT	31.8	31.3	30.8
TT	3.9	4.5	4.1
CFH rs800292
GG	52.6	59.8	56.9
GA	40.0	35.0	37.2
AA	5.3	5.3	6.0
ARMS2 (A69S) rs10490924
GG	62.6	60.7	62.0
GT	33.6	34.8	34.0
TT	3.8	4.5	4.0
C2/CFB (L9H) rs4151667
TT	90.8	90.7	91.1
TA/AA	9.2	9.3	8.9
C2/CFB (R32Q) rs641153
GG	84.4	83.7	81.9
GA/AA	15.6	16.3	18.1
C3 (R102G) rs2230199
CC	61.3	64.8	67.9
CG	33.9	31.2	28.8
GG	4.8	4.0	3.3
C3 rs433594
GG	39.4	39.4	35.1
GA	45.7	46.0	49.3
AA	14.9	14.6	15.6
CFI rs10033900
CC	27.4	25.6	27.6
CT	48.6	50.7	50.1
TT	24.0	23.7	22.4
LPL rs256
CC	73.1	72.6	72.3
CT/TT	26.9	27.4	27.7
LIPC rs12912415
AA	71.5	70.8	72.6
AG/GG	28.5	29.2	27.4
MYRIP rs2679798
AA	30.3	34.5	29.9
AG	49.7	47.5	48.9
GG	20.0	18.1	21.1
SKIV2L rs429608
GG	74.4	74.7	71.4
GA	23.6	23.1	26.7
AA	2.0	2.2	1.9
ABAC1 rs1883025
CC	54.4	56.3	55.5
CT	39.3	38.3	37.8
TT	6.3	5.4	6.7
CETP rs3764261
CC	46.4	44.5	45.3
CA	43.1	44.8	45.6
AA	10.5	10.7	9.1
TIMP3 rs5749482
GG	78.7	78.0	75.8
CG/CC	21.3	22.0	24.2
VEGFA rs943080
CC	24.9	26.6	26.6
TC	49.8	48.9	49.7
TT	25.3	24.4	23.8
COL8A1 rs13081855
GG	82.4	82.8	82.3
GT	16.9	16.3	16.9
TT	0.7	0.9	0.8
TNFRSF10A rs13278062
TT	28.0	26.3	27.0
GT	49.8	51.7	49.9
GG	22.2	21.9	23.1
FRK/COL10A1 rs3812111
TT	42.9	38.6	39.1
AT	44.8	47.3	46.9
AA	12.3	14.1	14.0
SLC16A8 rs8135665
CC	63.1	65.0	61.8
CT	32.7	31.3	33.7
TT	4.2	3.7	4.4
ADAMTS9 rs6795735
CC	33.9	33.5	36.6
TC	50.1	49.2	46.2
TT	15.9	17.3	17.2
TGFBR1 rs334353
TT	58.1	54.7	54.8
GT	36.1	39.0	38.8
GG	5.9	6.3	6.5
RAD51B rs8017304
AA	38.2	38.8	41.9
AG	46.6	47.9	44.1
GG	15.2	13.3	14.0
IER3/DDR1 rs3130783
AA	64.5	64.9	62.3
AG	32.0	26.0	33.4
GG	3.5	9.2	4.3
B3GALTL rs9542236
TT	31.0	31.9	33.1
CT	48.7	49.3	48.1
CC	20.3	18.7	18.8

Data are percentages unless otherwise indicated.

AMD = age-related macular degeneration; BDES = Beaver Dam Eye Study; BMES = Blue Mountain Eye Study; BMI = body mass index; RS = Rotterdam Study; SD = standard deviation.

Risk factors were tested separately for association with incident late AMD in all cohorts, adjusting for age and sex (Table 3, available at http://aaojournal.org). Most SNPs in the genes CFH, ARMS2, and CFHR5 were significant in all 3 cohorts (P < 0.05). The SNPs in the genes LIPC, TIMP3, ADAMTS9, IER3/DDR1, TNFRSF10A, TGFBR1, and B3GALTL were not significant in any single cohort. In all other genes, SNPs were significant in at least 1 of the cohorts. In all 3 studies, increasing severity levels of early AMD stages at baseline, based on the 3CC AMD severity scale (Table 1, available at http://aaojournal.org), were associated with a highly significant risk of incident late AMD. Of the environmental risk factors, current smoking showed a significant association in the RS and BMES, but BMI showed no significant association with incident late AMD in all 3 cohorts. To determine the best set of markers for each locus, all SNPs were analyzed per locus in a multivariate Cox proportional hazard analysis. A total of 26 SNPs were found to be suitable for further analyses.

Prediction models were built using various sets of risk factors and tested in the RS (Table 4). Model 1, a minimal model that included only age and sex, providedan AUC of 0.60 (95% confidence interval [CI], 0.55–0.65). Adding environmental and ocular factors improved the AUC to 0.78 (95% CI, 0.74–0.82, model 2). Adding only major AMD genes to model 1 increased the AUC to 0.73 (95% CI, 0.69–0.78, model 3). Next, a maximal genetic model was created that included all 26 SNPs. This increased the AUC to 0.82 (95% CI, 0.79–0.86, model 4). Finally, we combined all variables from models 1 to 4 to assess the best possible prediction. This resulted in an AUC of 0.88 (95% CI, 0.85–0.90, model 5). Validation of this model in the pooled dataset of the BDES and BMES (Table 5) showed an AUC of 0.85 (95% CI, 0.82–0.88). To further improve the prediction model, we pooled all 3 cohorts and re-estimated the risks of the variables included in model 5 (Table 6). The AUC in the 3 cohorts combined was 0.87 (95% CI, 0.85–0.89), and the model had a good calibration (P = 0.55). We also investigated the possibility of minimalizing this model. By using backward regression, 13 SNPs could be excluded from the model and provided a somewhat lower AUC of 0.86 (95% CI, 0.84–0.88) (Table 7, available at http://aaojournal.org). The model with the best AUC in the 3 cohorts combined dataset was used for further analyses.

Table 4. Predictive Values for the Tested Models in the Rotterdam Study.

Model	Variables	AUC (95% CI)	SE	No Late AMD/Late AMD
1	Minimal model	0.60 (0.55–0.65)	0.023	4621/132
	Age
	Sex
2	Nongenetic model	0.78 (0.74–0.82)	0.022	4561/132
	Age
	Sex
	AMD baseline grade
	Smoking
	BMI
3	Minimal genetic model	0.73 (0.69–0.78)	0.022	4226/121
	Age
	Sex
	ARMS2 rs10490924
	CFH rs1061170
	C2/CFB rs641153
	C3 rs2230199
	C2/CFB rs4151667
4	Maximal genetic model	0.82 (0.79–0.86)	0.016	4226/121
	Age
	Sex
	ARMS2 rs10490924
	CFH rs800292
	CFH rs12144939
	CETP rs3764261
	C2/CFB rs641153
	COL8A1 rs13081855
	C2/CFB rs4151667
	C3 rs433594
	TGBR1 rs334353
	SKIV2L rs429608
	C3 rs2230199
	VEGFA rs943080
	ADAMTS9 rs6795735
	CFH rs1061170
	TIMP3 rs5749482
	IER3/DDR1 rs3130783
	LPL rs256
	MYRIP rs2679798
	SLC16A8 rs8135665
	RAD51B rs8017304
	CFI rs10033900
	FRK/COL10A1 rs3812111
	ABCA1 rs1883025
	B3GALTL rs9542236
	LIPC rs12912415
	TNFRSF10A rs13278062
5	Maximal gene-environment model	0.88 (0.85–0.90)	0.015	4171/121
	Age
	Sex
	ARMS2 rs10490924
	CFH rs12144939
	CFH rs800292
	C3 rs433594
	C2/CFB rs641153
	TGBR1 rs334353
	SKIV2L rs429608
	CETP rs3764261
	C2/CFB rs4151667
	IER3/DDR1 rs3130783
	C3 rs2230199
	ADAMTS9 rs6795735
	LPL rs256
	COL8A1 rs13081855
	SLC16A8 rs8135665
	FRK/COL10A1 rs3812111
	CFH rs1061170
	TIMP3 rs5749482
	VEGFA rs943080
	MYRIP rs2679798
	CFI rs10033900
	TNFRSF10A rs13278062
	RAD51B rs8017304
	B3GALTL rs9542236
	ABCA1 rs1883025
	LIPC rs12912415
	AMD baseline grade
	Smoking
	BMI

AMD = age-related macular degeneration; AUC = area under the curve; BMI = body mass index; CI = confidence interval; SE = standard error.

Table 5. Model 5, the Maximal Gene-Environment Model, in the Three Cohorts.

Study	AUC (95% CI)	SE	No Late AMD/Late AMD
RS	0.88 (0.85–0.90)	0.015	4171/121
BDES + BMES pooled	0.85 (0.82–088)	0.015	3634/156
3CC*	0.87 (0.85–0.89)	0.010	7805/277

3CC = Three Continent AMD Consortium; AMD = age-related macular degeneration; AUC = area under the curve; BDES = Beaver Dam Eye Study; BMES = Blue Mountain Eye Study; CI = confidence interval, RS = Rotterdam Study; SE = standard error.

^*Based on risk estimates reanalyzed in the complete data set.

Table 6. Risk Estimates from Cox Proportional Hazard Analysis Based on the Three Continent AMD Consortium.

Variable	Code	Beta Coefficient per Code
Age	<65 = 0/65–75 = 1/75+ = 2	0/1.558/2.433
Sex	M = 0/F = 1	0/0.320
ARMS2 rs10490924	GG = 0/GT = 1/TT = 2	0/0.779/1.720
CFH rs800292	GG = 0/GA = 1/AA = 2	0/−0.899/−1.614
C2/CFB rs4151667	TT = 0/TA or AA = 1	0/−1.245
CFH rs12144939	GG = 0/GT = 1/TT = 2	0/−0.947/−1.195
COL8A1 rs13081855	GG = 0/GT = 1/TT = 2	0/0.223/0.890
C3 rs2230199	CC = 0/GC = 1/GG = 2	0/−0.033/0.755
SLC16A8 rs8135665	CC = 0/TC = 1/TT = 2	0/0.313/0.648
C3 rs433594	GG = 0/GA = 1/AA = 2	0/−0.110/−0.591
C2/CFB rs641153	GG = 0/GA or AA = 1	0/−0.592
SKIV2L rs429608	GG = 0/GA = 1/AA = 2	0/0.027/0.590
CETP rs3764261	CC = 0/CA = 1/AA = 2	0/0.215/0.478
ADAMTS9 rs6795735	CC = 0/TC = 1/TT = 2	0/0.130/0.424
RAD51B rs8017304	AA = 0/AG = 1/GG = 2	0/−0.414/−0.138
TIMP3 rs5749482	GG = 0/GC or CC = 1	0/−0.357
TGBR1 rs334353	TT = 0/TG = 1/GG = 2	0/0.039/−0.336
CFH rs1061170	TT = 0/TC = 1/CC = 2	0/0.175/0.278
FRK/COL10A1 rs3812111	TT = 0/TA = 1/AA = 2	0/−0.278/−0.118
CFI rs10033900	CC = 0/TC = 1/TT = 2	0/−0.070/−0.223
TNFRSF10A rs13278062	TT = 0/TG = 1/GG = 2	0/0.093/0.196
B3GALTL rs9542236	TT = 0/TC = 1/CC = 2	0/−0.231/−0.169
IER3/DDR1 rs3130783	AA = 0/AG = 1/GG = 2	0/0.029/0.166
MYRIP rs2679798	AA = 0/AG = 1/GG = 2	0/0.059/0.156
VEGFA rs943080	CC = 0/TC = 1/TT = 2	0/0/0.098
ABCA1 rs1883025	CC = 0/TC = 1/TT = 2	0/−0.046/0.076
LIPC rs12912415	AA = 0/AG or GG = 1	0/−0.098
LPL rs256	CC = 0/TC or TT = 1	0/−0.048
AMD baseline grade	Level 10 = 0/Level 20 = 1/Level 30 = 2/Level 40 = 3	0/1.458/2.560/3.398
Smoking	Never = 0/Past = 1/Current = 2	0/0.164/0.651
BMI (kg/m2)	<25 = 0/25+ = 1	0/0.007

AMD = age-related macular degeneration; BMI = body mass index; F = female; M = male.

We calculated a risk score (Table 6) based on the beta coefficients from the pooled analysis, which ranged from −3.99 to 7.56. We rounded off the estimates and plotted their distribution stratified for incident or no incident late AMD (Fig 2). The plot showed a bimodal distribution with a large frequency difference between the groups for scores <2 and scores >3. The frequencies of risk scores 2 and 3 showed relatively small differences between cases and noncases. Of note, all participants (n = 8) with risk score 8 had risk alleles in the CFH, ARMS2, and C3 genes and no protective alleles in C2/CFB. In contrast, all those with risk score −3 (n = 29) carried a protective variant in C2/CFB and were free of variants in CFH and ARMS2, except 1 person who carried a heterozygous variant in ARMS2. Risk of incident late AMD for individuals with a risk score of 6 to 8 showed a hazard ratio (HR) of 23.2 (95% CI, 15.9–34.0). Risk of incident late AMD for individuals with a risk score of −3 to 0 showed an HR of 0.02 (95% CI, 0.01–0.04).

Figure 2.

Distribution of risk score in incident late age-related macular degeneration (AMD) and no late AMD in the Three Continent AMD Consortium, and hazard ratio (HRs) of incident late AMD. The x-axis represents the risk score category, the left y-axis represents frequency as percentages, and the right y-axis represents the HR of incident late AMD. The dark grey bars represent no late AMD, and the light grey bar represents incident late AMD. The black line represents the HR of incident late AMD. Category 3 is the reference category (R) and has an HR of 1.00. Risk scores −3 to 0 and 6 to 8 were combined for HR. Error bars indicate the 95% confidence interval of the HRs.

Cumulative risk of incident late AMD was calculated for each risk score and compared with the overall cumulative risk of incident late AMD. Individuals (n = 181) with a risk score 6 to 8 had a cumulative risk of 65.6% (standard error [SE], 0.057) to develop late AMD at age 90 years, and those (n = 2751) with a risk score of –3 to 0 had virtually no risk of developing incident late AMD (0.5%; SE, 0.002). The overall risk of AMD for our study population before testing was 17.4% (SE, 0.013) at age 90 years (Fig 3), which was significantly different from all strata apart from risk score 3 (P = 0.71).

Figure 3.

Cumulative risk of incident late age-related macular degeneration (AMD) prior to and after testing of prediction model in the Three Continent AMD Consortium (3CC). A, Overall cumulative risk of incident late AMD in 3CC; the x-axis represents the age of onset for incident late AMD cases and age at last examination for noncases; y-axis represents the cumulative risk of incident late AMD. The number at risk for each decade is presented underneath the figure. B, Cumulative risk of incident late AMD stratified for risk strata 0–6 in 3CC; x-axis represents the age of onset for incident late AMD cases and age at last examination for noncases; y-axis represents the cumulative risk of incident late AMD. The risk scores are depicted by various colors. The number at risk for each decade and each risk score is presented underneath the figure.

Discussion

In 3 independent population-based studies from 3 continents, we investigated all well-known genetic and nongenetic risk factors for AMD. We found that the best prediction for late AMD was based on age, sex, 26 genetic variants, 2 environmental variables, and early AMD phenotype. The accuracy of a prediction model including all these variables was 0.88 in the RS. Because similar risk estimates were found in the BDES and BMES, the model proved to be well generalizable to people of Caucasian descent living on other continents. Translation of the model to the individual level provided good discrimination between those with a high lifetime risk of late AMD and those with virtually no risk, a risk difference of 65%.

A major strength of our study is the inclusion of a general population unbiased by AMD risk factors. The study samples had included a wide spectrum of AMD lesion phenotypes but not only the extreme ends of disease, which is representative of real scenarios in the population. Inclusion of a wide spectrum of risk factor distributions and various levels of risk profiles of our population-based samples ensures a realistic, less biased prediction for all risk categories. Additional strengths are the use of longitudinal observational samples to predict incident cases and validation in 2 independent population-based cohorts with similar study designs. All the strengths facilitate our comprehensive analyses and calculation of cumulative risk of incident late AMD.

Limitations included the relatively low number of incident late AMD cases (n = 363), hampering further risk estimation to AMD subtypes. In addition, we did not include several risk factors, such as dietary factors, biomarkers, or rare genetic variants.^10,40–43 Dietary factors and biomarkers are difficult variables to obtain, but their inclusion would have improved the sensitivity of the predictive value. Inclusion of genetic mutations is unlikely to contribute to population risk because of their low frequencies. Finally, the 3 cohorts had subtle differences in methodology, which have been discussed by Klein et al.³⁷

Most previously published AMD prediction models have been based on case-control studies (Table 8).^12–28 Most of these models included demographic, genetic, and environmental factors and reached a good prediction for AMD (AUC, 0.68−0.94). The study that reported the highest AUC (0.94) included complement activation.²² However, measurement of activation fragments requires rather intense workup and specific expertise and is therefore unlikely to occur in a standard clinical setting. The other studies have some drawbacks as well. Gold et al¹⁴ reported a sensitivity of 70% and a specificity of 50% for their model, making the prediction of low-risk outcomes inaccurate. Hageman et al¹⁶ showed a better specificity and sensitivity, but their model did not incorporate any nongenetic factors, including age.¹⁶ Furthermore, their model was based on prevalence data, which is less appropriate for estimation of prognosis.⁴⁴ Most other case-control studies also lacked follow-up data. The reports with follow-up data were almost inclusively based on data from the Age-Related Eye Disease Study.^20,23,24,27 Although the investigated people are from the same source population, the prediction studies using this population differed substantially from each other in design and inclusion of risk factors, leading to great variation in prediction outcomes. Our study shows that the sensitivity of risk prediction depends on the number of variables included in the model, and highest sensitivity is achieved with a full model including the major genes, many of the recently discovered minor genetic variants, smoking, BMI, and existing AMD phenotypes.

Table 8. Prediction Models in the Literature.

Author	Year	Study Population/ Design	Noncases/ Cases	Follow-up	Prediction Model Variables				AUC	Sensitivity	Specificity	Validation in Independent Dataset?	Remarks
					Demographic	Environmental	Genetic	Ocular
Gold et al14	2006	Case-control	114/350				CFH, C2/CB			0.74	0.56	Yes
Hughes et al17	2007	Case-control	266/401			Smoking	CFH, ARMS2					No	Risk score only CNV case
Jakobsdottir et al18	2008	Case-control and family cohort	168/187; 1095/429		Age, sex	Smoking	CFH, ARMS2, C2/FB			0.70	0.74	No
Jakobsdottir et al19	2009	Case-control and family cohort	168/187; 1095/429				CFH, ARMS2, C2/FB		0.79			No
Seddon et al23	2009	AREDS, RCT	1167/279	6.3 yrs	Age, sex	Education, smoking, BMI, antioxidant/zinc or both	CFH, ARMS2, C2/FB, C3	Baseline grade	0.822	0.83	0.68	No
Reynolds et al22	2009	Case-control	60/110		Age, sex	Smoking, BMI, action fragments	CFH, ARMS2, C2/FB, C3, CFI		0.94			No
Gibson et al13	2010	Case-control	470/470		Age, sex	Smoking	CFH, ARMS2, C3, SERPING1		0.83	0.76	0.76	No
McKay et al21	2010	Case-control	436/437		Age	Smoking	CFH, ARMS2, C2/FB, C3		0.86			No
Zanke et al28	2010	Risks from literature	NA			Smoking	CFH, ARMS2, C3, mtA4917G					No
Chen et al12	2011	AREDS (723), case-control (1121)	509/1335		Age, sex	Smoking, BMI	CFH, ARMS2, C2/FB, C3		0.82	0.755	0.747	No
Hageman et al16	2011	Case-control	467/482				CFH, ARMS2, C2/FB, C3, CFHR4, CFHR5, F13B		0.80	0.82	0.63	Yes	Only CNV cases
Klein et al20	2011	AREDS, RCT	2034/688	9.3 yrs	Age	Family history, smoking	CFH, ARMS2	Baseline grade, drusen presence, advanced AMD in other eye	0.872†			Yes
Seddon et al24	2011	AREDS, RCT	2118/819	9.2 yrs	Age, sex	Education, smoking, BMI, antioxidant	CFH, ARMS2, C2/FB, C3	Drusen size, advanced AMD in other eye	0.908* 0.876†			No
Spencer et al25	2011	Case-control	216/349		Age	Smoking	CFH, ARMS2, C2/FB, C3		0.84	0.770	0.741	Yes	Outcome is early and late AMD
Spencer et al25	2011	Nested case-control	148/85		Age	Smoking	CFH, ARMS2, C2/FB, C3			0.633	0.707	Yes	Outcome is early and late AMD
Ying et al26	2011	CAPT, trial	618/324	5 yrs	Age	Smoking, hypertension	…	Baseline grade, night vision score	0.68			No	Only GA cases
Grassmann et al15	2012	Case-control	786/986				CFH, ARMS2, C2/FB, LIPC, APOE, PLA2G12A, TIMP3		0.813			No
Yu et al27	2012	AREDS, RCT	1112/1448	10.3 yrs	Age, sex	Education, smoking, BMI, antioxidant	CFH, ARMS2, C2/FB, C3, CFI, LIPC, CFI, LIPC, TIMP3, CETP ABCA1, COL8A1, APOE	Fellow eye status	0.895* 0.883†			No	Outcome is early and late AMD

AMD = age-related macular degeneration; AREDS = Age-Related Eye Disease Study; AUC = area under the curve; BMI = body mass index; randomized controlled trial.

^*Based on 10-year follow-up.

^†Based on 5-year follow-up.

What may be the benefits of prediction tests? Most current counseling provided to family members of late AMD cases is based on clinical parameters. A prediction test may improve the identification of true high-risk individuals. Because the estimation of cumulative risk of incident AMD makes the risk apparent, it may encourage individuals to alter their lifestyle with the aim of decreasing the risk of AMD. For instance, one can stop smoking, eat foods rich in antioxidants, and increase physical exercise to lower risk of progression to late AMD.^9,45 There may be benefits for patients with late AMD. Various studies have shown that patients with neovascular AMD who do not respond to anti-VEGF therapy are at higher genetic risk.^46–29 These patients may need more intensive treatment regimens. Last, current intervention trials select study participants mainly on the basis of phenotypes. Inclusion of high-risk individuals, identified by a prediction test, may improve homogeneity of the study population and prediction of AMD outcome events.

In conclusion, a risk score based on a large number of genetic risk variants for AMD, the environmental factors smoking and BMI, and early AMD phenotype provided a good prediction of incident late AMD cases in this study. A model incorporating nongenetic factors performed better than a model based on a minimal number of genetic factors, but after inclusion of many genes the model performed better than a model including only nongenetic factors. Inclusion of all risk factors provided the best prediction. Because personalized medicine is the future, prediction tests will become more thoroughly implemented as clinical tools. In such case, only comprehensive tests will be useful for AMD.